Model training method based on edge-end collaboration and related equipment

Abstract

The invention provides a model training method based on edge-end cooperation and related equipment, and the method comprises the steps: carrying out the splitting of a to-be-trained model according to the resource parameters of mobile equipment, and obtaining a first model and a second model; performing knowledge distillation on the second model based on local data of the mobile equipment to obtain a distillation model; sending the first parameter of the distillation model to a data processor to obtain a second parameter, and uploading the second parameter to a server; calculating a fourth parameter based on the second parameter and a third parameter of the to-be-trained model; updating the first model based on the fourth parameter to obtain a third model and a fifth parameter of the third model; based on the fifth parameter, training the distillation model to obtain a sixth parameter of a fourth model; and when training is stopped, taking the third model and the fourth model as trained models. According to the method and the device, the federated learning and the split learning architecture are combined, so that model training is promoted under the conditions of limited computing power resources and relatively high data privacy requirements.

Description

Technical Field

[0001] The present application relates to the field of model training technology, and in particular to a model training method and related equipment based on edge-end collaboration. Background Art

[0002] With the continuous development and widespread application of large language model technology, the number of corresponding model parameters has grown rapidly like an exponential function curve. The rapid iteration of technology in the field of large language models and the increasing complexity of models are facing the growing demand for user data privacy protection and data silos. Large language models face huge challenges in obtaining large-scale high-quality training data. Traditional large language model training is to upload data to the cloud, and the model and data are decentralized after training in the cloud, relying on the computing power and large storage space provided by the cloud. Summary of the invention

[0003] In view of this, the purpose of this application is to propose a model training method and related equipment based on edge collaboration.

[0004] Based on the above objectives, the present application provides a model training method based on edge-end collaboration, including:

[0005] Splitting the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model;

[0006] Based on local data of the mobile device, using the mobile device to perform knowledge distillation on the second model to obtain a distilled model;

[0007] Using a mobile device to send a first parameter of the distillation model to a data processor to obtain a second parameter, and using the data processor to upload the second parameter to a server;

[0008] Using the server, based on the second parameter and the third parameter of the model to be trained, a fourth parameter is calculated;

[0009] Based on the fourth parameter, using the server to update the first model to obtain a third model and a fifth parameter of the third model;

[0010] Based on the fifth parameter, using the mobile device to train the distillation model to obtain a sixth parameter of the fourth model;

[0011] In response to reaching a condition for stopping training, the third model and the fourth model are used as trained models.

[0012] In a possible implementation, the mobile device resource parameters include hardware resource parameters, delay sensitivity, load balancing, and model structure;

[0013] The method of splitting the model to be trained according to the obtained mobile device resource parameters to obtain the first model and the second model includes:

[0014] The model to be trained is split based on the hardware resource parameters, delay sensitivity, load balancing and model structure to obtain the first model and the second model.

[0015] In a possible implementation, the step of performing knowledge distillation on the second model based on the local data of the mobile device and using the mobile device to obtain the distilled model includes:

[0016] Using the server to train the second model to obtain a fifth model;

[0017] Based on the local data and the learning loss of the soft label distribution of the fifth model, the second model is subjected to knowledge distillation using the mobile device to obtain the distilled model.

[0018] In a possible implementation, the step of sending the first parameter of the distillation model to a data processor using a mobile device to obtain a second parameter, and uploading the second parameter to a server using the data processor includes:

[0019] Sending the first parameter of the distillation model to the data processor using a mobile device;

[0020] Using the data processor to perform aggregation processing on the first parameter to obtain the second parameter;

[0021] The second parameter is uploaded to the server using the data processor.

[0022] In a possible implementation manner, the calculating the fourth parameter based on the second parameter and the third parameter of the model to be trained includes:

[0023]

[0024] Among them, w new Indicates the fourth parameter, w global Represents the third parameter, represents the second parameter, i represents the i-th switch, η represents the learning rate of the model to be trained, and N represents the number of mobile devices participating in this round of training.

[0025] In a possible implementation manner, the updating the first model by using a server based on the fourth parameter to obtain a third model and a fifth parameter of the third model includes:

[0026] The fourth parameter is used as a parameter of the first model, and the first model is updated by using a server to obtain a third model and a fifth parameter of the third model.

[0027] In a possible implementation, the sixth parameter is calculated by the following formula:

[0028]

[0029] Wherein, r″ represents the sixth parameter of the fourth model, a represents the update ratio, r represents the seventh parameter obtained by the server through training the model to be trained, and r′ represents the first parameter.

[0030] Based on the same inventive concept, the embodiment of the present application also provides a model training device based on edge-end collaboration, including:

[0031] A splitting module is configured to split the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model;

[0032] The mobile device is configured to perform knowledge distillation on the second model based on local data of the mobile device to obtain a distilled model;

[0033] A mobile device configured to send the first parameter of the distillation model to a data processor to obtain a second parameter, and upload the second parameter to a server using the data processor;

[0034] The server is configured to calculate a fourth parameter based on the second parameter and a third parameter of the model to be trained;

[0035] The server is configured to obtain a third model and a fifth parameter of the third model by updating the first model based on the fourth parameter;

[0036] A mobile device, configured to train the distillation model based on the fifth parameter to obtain a sixth parameter of the fourth model;

[0037] The stopping module is configured to use the third model and the fourth model as trained models in response to reaching a condition for stopping training.

[0038] Based on the same inventive concept, an embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the model training method based on edge collaboration as described in any one of the above items.

[0039] Based on the same inventive concept, an embodiment of the present application also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute any of the above-mentioned edge-to-edge collaboration-based model training methods.

[0040] From the above, it can be seen that the model training method and related equipment based on edge collaboration provided by the present application, according to the obtained mobile device resource parameters, splits the model to be trained to obtain a first model and a second model; based on the local data of the mobile device, uses the mobile device to perform knowledge distillation on the second model to obtain a distillation model; uses the mobile device to send the first parameter of the distillation model to the data processor to obtain the second parameter, and uses the data processor to upload the second parameter to the server; uses the server to calculate the fourth parameter based on the second parameter and the third parameter of the model to be trained; based on the fourth parameter, uses the server to update the first model to obtain the third model and the fifth parameter of the third model; based on the fifth parameter, uses the mobile device to train the distillation model to obtain the sixth parameter of the fourth model; in response to reaching the condition of stopping training, the third model and the fourth model are used as trained models. The present application embodiment of the present invention first constructs a large language model collaborative training framework based on federated learning and split learning. Under this framework, the participants do not need to exchange original data. Through the federated learning mechanism, the model parameters can be trained locally, and the encrypted model updates can be uploaded to the central server, ensuring data privacy and compliance. Taking into account the large number of parameter exchange processes in federated learning, an in-network aggregation method based on a data processor (Data Processing Unit, DPU) is designed. The DPU is deployed in the switch to aggregate parameters while transmitting data, reduce network load, and increase overall throughput. At the same time, using the split learning strategy, the large language model is divided into smaller sub-models, which are deployed on different nodes for parallel training, and then the model updates are integrated through an encryption protocol, which not only significantly improves the training speed, but also reduces the computational burden of a single participant. Secondly, the present invention designs a large language model fine-tuning method based on knowledge distillation. In the training of large language models, knowledge distillation can help this article transfer the knowledge of large models to small models, thereby realizing the lightweight and efficient model, and reducing the pressure that federated learning brings to the network transmission environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present application or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 This is a flow chart of a model training method based on edge-end collaboration according to an embodiment of the present application;

[0043] Figure 2 A schematic diagram of the overall architecture of the model training method based on edge-end collaboration according to an embodiment of the present application;

[0044] Figure 3 A schematic diagram of model splitting evaluation data in an embodiment of the present application;

[0045] Figure 4 This is a schematic diagram of the model splitting of an embodiment of the present application;

[0046] Figure 5 A schematic diagram of the knowledge distillation steps of an embodiment of the present application;

[0047] Figure 6 A schematic diagram showing the visualization of the knowledge distillation training results of the split model in an embodiment of the present application;

[0048] Figure 7 This is a schematic diagram of the structure of a model training device based on edge-end collaboration according to an embodiment of the present application;

[0049] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0051] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0052] It is understandable that before using the technical solutions of each embodiment of the present disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved will be informed to the user in an appropriate manner, and the user's authorization will be obtained.

[0053] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly remind the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can independently choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0054] As an optional but non-limiting implementation, in response to receiving the user's active request, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0055] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0056] As mentioned in the background technology section, with the continuous development and widespread application of large language model technology, the number of corresponding model parameters has grown rapidly like an exponential function curve. The rapid iteration speed of technology in the field of large language models and the increasing complexity of models are facing the growing demand for user data privacy protection and data silos. Large language models face huge challenges in obtaining large-scale high-quality training data. Traditional large language model training is to upload data to the cloud, and the model and data are decentralized after training in the cloud, relying on the computing power and large storage space provided by the cloud.

[0057] Based on the above considerations, the embodiment of the present application proposes a model training method based on edge collaboration, according to the obtained mobile device resource parameters, the model to be trained is split to obtain a first model and a second model; based on the local data of the mobile device, the second model is distilled using the mobile device to obtain a distillation model; the first parameter of the distillation model is sent to the data processor using the mobile device to obtain the second parameter, and the second parameter is uploaded to the server using the data processor; the server is used to calculate the fourth parameter based on the second parameter and the third parameter of the model to be trained; based on the fourth parameter, the first model is updated using the server to obtain the third model and the fifth parameter of the third model; based on the fifth parameter, the distillation model is trained using the mobile device to obtain the sixth parameter of the fourth model; in response to the condition of stopping training, the third model and the fourth model are used as trained models. The embodiment of the present application The present invention first constructs a large language model collaborative training framework based on federated learning and split learning. Under this framework, the participants do not need to exchange original data. Through the federated learning mechanism, the model parameters can be trained locally, and the encrypted model updates can be uploaded to the central server, ensuring data privacy and compliance. Taking into account the large number of parameter exchange processes in federated learning, an in-network aggregation method based on the data processor DPU is designed. The DPU is deployed in the switch to aggregate parameters while transmitting data, reduce network load, and increase overall throughput. At the same time, using the split learning strategy, the large language model is divided into smaller sub-models, which are deployed on different nodes for parallel training, and then the model updates are integrated through encryption protocols, which not only significantly improves the training speed, but also reduces the computational burden of individual participants. Secondly, the present invention designs a large language model fine-tuning method based on knowledge distillation. In the training of large language models, knowledge distillation can help this article transfer the knowledge of large models to small models, thereby realizing the lightweight and efficient model, and reducing the pressure of federated learning on the network transmission environment.

[0058] The technical solutions of the embodiments of the present application are described in detail below through specific examples.

[0059] refer to Figure 1 The model training method based on edge-end collaboration in the embodiment of the present application includes the following steps:

[0060] Step S101, splitting the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model;

[0061] Step S102, based on the local data of the mobile device, using the mobile device to perform knowledge distillation on the second model to obtain a distilled model;

[0062] Step S103, using a mobile device to send the first parameter of the distillation model to a data processor to obtain a second parameter, and using the data processor to upload the second parameter to a server;

[0063] Step S104, using the server to calculate a fourth parameter based on the second parameter and the third parameter of the model to be trained;

[0064] Step S105, based on the fourth parameter, using the server to update the first model to obtain a third model and a fifth parameter of the third model;

[0065] Step S106, based on the fifth parameter, using the mobile device to train the distillation model to obtain a sixth parameter of the fourth model;

[0066] Step S107, in response to reaching a condition for stopping training, taking the third model and the fourth model as trained models.

[0067] This application considers combining the technical features of federated learning and split learning in the process of large language model training, and can design a more flexible and secure large language model training system for edge environments with limited computing hardware resources and network transmission resources.

[0068] refer to Figure 2 , which is a schematic diagram of the overall architecture of the model training method based on edge collaboration in an embodiment of the present application.

[0069] like Figure 2 As shown in the overall architecture diagram, in the embodiment of the present application, for a specific large language model, taking the Bert model as an example, the training steps are roughly divided into the following six steps.

[0070] Step 1: Consider mobile device resources, computing latency, and model accuracy to determine the cutting layer of the large language model. Split the large language model into two sub-models, called the server-side model and the device-side model. All participating devices download the latest device-side model from the server.

[0071] Step 2: Each participating device uses local data to perform knowledge distillation on the device-side model to obtain the device-side model parameters and related calculation gradients. The knowledge distillation technology is first used on the device side to optimize the small model so that it can imitate the behavior of the teacher model as much as possible. In theory, if all devices use knowledge distillation and the distillation effect is good, the model aggregation result should be able to reflect some of the knowledge of the teacher model. However, since information loss is inevitable during the distillation process, the performance of the final aggregated model will be slightly lower than the effect of directly training the complete model on the device side.

[0072] Step 3: After each device independently completes the training and distillation of the device model, the updated model parameter Δw i (First parameter) sent to the server.

[0073] Step 4: Deploy the DPU in the switch and aggregate the parameters of each device in the switch, that is:

[0074]

[0075] in is the second parameter, i represents the i-th switch, w j is the model parameter of the jth device passing through the switch.

[0076] Step 5: The server completes the model parameter aggregation based on the gradient. The server receives all the device parameters from the switch and aggregates these updates using a simple average method:

[0077]

[0078] Among them, w new Indicates the fourth parameter, w golbla Represents the third parameter, represents the second parameter, i represents the i-th switch, η represents the learning rate of the model to be trained, and N represents the number of mobile devices participating in this round of training.

[0079] Step 6: Updated global model parameters w new , as the input of server model training, the server-side model training is completed on the server side, and the new model parameters are transmitted to the device-side model for updating. The update strategy adopts:

[0080]

[0081] Wherein, r″ represents the sixth parameter of the fourth model, a represents the update ratio, r represents the seventh parameter obtained by the server through training the model to be trained, and r′ represents the first parameter.

[0082] Repeat steps 2-6 above until the model results converge.

[0083] Next, each step is described in detail. For the above step S101, the model to be trained is split according to the obtained mobile device resource parameters to obtain a first model and a second model.

[0084] In some embodiments, the mobile device resource parameters include hardware resource parameters, delay sensitivity, load balancing and model structure; the model to be trained is split according to the acquired mobile device resource parameters to obtain a first model and a second model, including: based on the hardware resource parameters, delay sensitivity, load balancing and model structure, the model to be trained is split to obtain the first model and the second model.

[0085] In this embodiment, the first model is a server-side model, and the second model is a device-side model. For the model splitting part, taking the large language model: semantic representation model (Bidirectional Encoder Representations from Transformers, BERT) as an example, the large model splitting process based on the Transformer structure can be combined with the split learning algorithm to perform model vertical splitting. Taking the BERT model as an example, for the BERT-Base model, under standard hardware configuration and optimization conditions, with a smaller batch size (16) and a medium-length sentence (128 tokens), a single forward propagation takes about 2 seconds. A one-way propagation experiment of the model was conducted.

[0086] refer to Figure 3 , which is a schematic diagram of the model splitting evaluation data of an embodiment of the present application.

[0087] refer to Figure 4 , which is a schematic diagram of the model splitting of an embodiment of the present application.

[0088] from Figure 3 It can be seen that the expected training delay for each cutting layer is basically the same. Since there is no difference between each encoder module of Bert, the enumeration strategy is used to split the BERT model vertically, such as Figure 4 As shown, the specific layers of segmentation in the BERT model are enumerated.

[0089] In addition, when dividing the model into device-side model and server-side model, the following key factors should usually be considered:

[0090] Hardware resource allocation: computing power, memory size, and bandwidth limitations on the device side (such as mobile phones and embedded devices). High-performance computing resources on the server side can handle more complex calculations and larger model capacity.

[0091] Latency sensitivity: The device-side model should minimize latency to achieve real-time or low-latency response, which requires that part of the model is lightweight enough. Although the server-side model does not need to consider real-time performance too much, the overall communication and computing latency must still be within an acceptable range.

[0092] Computational load balancing: Considering the computational intensity, the parts with higher computational complexity are placed on the server side, and simple pre-processing or preliminary feature extraction are placed on the device side.

[0093] Model structure characteristics: The choice of cutting layer should try not to affect the performance of the model. Usually, it is chosen to cut at the middle layer with a higher level of abstraction. This can ensure that the front-end device can efficiently extract basic features, while allowing the server to focus on understanding and making decisions on high-level features.

[0094] Communication cost: When designing model splits, the communication overhead between devices and servers needs to be considered. Smaller intermediate activation vectors can reduce data transmission time.

[0095] Security and privacy protection: When considering model splitting, user privacy must also be taken into account. It is necessary to ensure that sensitive information is processed on the device side and avoid uploading it to the server.

[0096] Model validity: The selection of cutting layers should be based on the principle of minimizing the performance loss of the model. Experiments can be used to verify the impact of different cutting strategies on model accuracy.

[0097] Therefore, when determining the large language splitting strategy, it is necessary to weigh multiple factors such as large language model performance, hardware resources, communication efficiency, privacy protection, etc., and find the optimal model structure split point so that the entire system can operate effectively under the distributed architecture. The selection of each parameter has its own particularity and is not a technical feature that can be easily thought of by technicians in this field.

[0098] With respect to step S102, knowledge distillation of the second model is performed using the mobile device based on local data of the mobile device to obtain a distilled model.

[0099] In some embodiments, the method of performing knowledge distillation on the second model based on the local data of the mobile device and obtaining the distilled model using the mobile device includes: training the second model using a server to obtain a fifth model; and performing knowledge distillation on the second model based on the learning loss of the local data and the soft label distribution of the fifth model using the mobile device to obtain the distilled model.

[0100] refer to Figure 5 , which is a schematic diagram of the knowledge distillation steps of an embodiment of the present application.

[0101] In this embodiment, if Figure 5The device-side model optimization strategy based on knowledge distillation shown is mainly a model compression and transfer learning technology, which allows knowledge to be extracted from a large and complex teacher model (usually an over-parameterized model) and effectively transferred to a small student model, so that the student model can reduce computing resource consumption and improve inference speed while maintaining high performance.

[0102] First, you need to get the pre-trained device-side model (fifth model) from the server side and use it as the teacher model.

[0103] The teacher model is then used to generate soft labels. In this embodiment, the teacher model can not only output hard category labels, but also output the probability distribution of each category (i.e., the aforementioned "soft labels").

[0104] In the embodiment, the distillation model is obtained by calculating the loss function and using the gradient descent method to train. The knowledge distillation loss function expression is as follows. The first part of the loss function refers to the cross loss of the softmax output and the soft label of the student model under the same temperature T, that is:

[0105]

[0106] Among them, z i Represents the logical value of the student model output, T represents the hyperparameter in the knowledge distillation process - temperature, which can control the smoothness of the probability distribution of soft labels, thereby affecting the learning process of the student model. K represents the total number of labels. Similarly, the second part of the loss function is the cross loss between the output of softmax and the predicted value and the true value when the temperature T = 1, that is:

[0107]

[0108] The student model (second model) in knowledge distillation is trained on the same task, but its loss function not only contains the conventional cross entropy loss, but also adds the learning loss of the soft label distribution provided by the teacher model. The student model attempts to imitate the decision-making behavior of the teacher model on the entire sample space, thereby absorbing the knowledge of the teacher model. It should be noted that in the overall embodiment, the second model is the device-side model, but in the process of knowledge distillation, the second model is the student model, and the two do not conflict.

[0109] In addition to knowledge transfer at the output level, knowledge distillation can also be achieved through the similarity of feature mappings between layers, that is, making the intermediate layer features of the student model as close as possible to the corresponding layer features of the teacher model.

[0110] According to the characteristics of the student model, the model structure can be further optimized, such as pruning and quantization, to reduce the complexity of the model while ensuring the distillation effect.

[0111] The above-mentioned device-side model optimization strategy based on knowledge distillation can not only be used to miniaturize the model, but also to accelerate the training and reasoning process, which is of great significance for natural language processing deployed in resource-constrained environments. At the same time, this strategy is also applicable to cross-domain transfer learning and incremental learning scenarios, which helps to improve the generalization ability and adaptability of the model in new environments. Only performing knowledge distillation on the device-side model will not directly affect the process of model aggregation, but it will have a positive impact on the following aspects:

[0112] The performance of the device-side model after distillation is expected to be close to or reach the level of the teacher model, so that even during the network aggregation process, the prediction quality of a single node can be improved, and the accuracy of the overall aggregation results will also benefit.

[0113] Generally speaking, smaller models mean lower memory usage and faster inference speed, which can potentially reduce the communication cost of transferring data or model parameters between devices. Especially in federated learning or collaborative training, the compactness of the on-device model helps optimize the efficiency of the entire system.

[0114] For edge devices with limited resources, the distilled small models are more suitable for deployment, reduce energy consumption, and adapt to more hardware conditions, which enables the effective participation of more devices in large-scale distributed environments.

[0115] With respect to step S103, a first parameter of the distillation model is sent to a data processor using a mobile device to obtain a second parameter, and the second parameter is uploaded to a server using the data processor.

[0116] In some embodiments, using a mobile device to send the first parameter of the distillation model to a data processor to obtain a second parameter, and using the data processor to upload the second parameter to a server includes: using a mobile device to send the first parameter of the distillation model to the data processor; using the data processor to aggregate the first parameter to obtain the second parameter; and using the data processor to upload the second parameter to the server.

[0117] In this embodiment, as described in the above steps 3 and 4:

[0118] After each device independently completes the training and distillation of the device model, the updated model parameter Δw i (First parameter) sent to the server.

[0119] Deploy the DPU in the switch and aggregate the parameters of each device in the switch, namely:

[0120]

[0121] in is the second parameter, i represents the i-th switch, w j is the model parameter of the jth device passing through the switch.

[0122] Further, with respect to step S104, a fourth parameter is calculated using the server based on the second parameter and the third parameter of the model to be trained.

[0123] In some embodiments, the calculating of the fourth parameter based on the third parameter of the model to be trained by the second parameter includes:

[0124]

[0125] Among them, w new Indicates the fourth parameter, w global Represents the third parameter, represents the second parameter, i represents the i-th switch, η represents the learning rate of the model to be trained, and N represents the number of mobile devices participating in this round of training.

[0126] In this embodiment, the server side completes the model parameter aggregation according to the gradient. The server receives all the device-side parameters from the switch and uses a simple averaging method to aggregate these updates.

[0127] With respect to step S105, based on the fourth parameter, the first model is updated by using the server to obtain a third model and a fifth parameter of the third model.

[0128] In some embodiments, based on the fourth parameter, the first model is updated using a server to obtain a third model and a fifth parameter of the third model, including: taking the fourth parameter as a parameter of the first model, updating the first model using a server to obtain a third model and a fifth parameter of the third model.

[0129] In this embodiment, the updated global model parameter w new , as the input of server model training, the server-side model training is completed on the server side, and new model parameters (fifth parameters) are obtained.

[0130] Further, with respect to step S106, based on the fifth parameter, the distillation model is trained using the mobile device to obtain a sixth parameter of the fourth model.

[0131] In some embodiments, the sixth parameter is calculated by the following formula:

[0132]

[0133] Wherein, r″ represents the sixth parameter of the fourth model, a represents the update ratio, r represents the seventh parameter obtained by the server through training the model to be trained, and r′ represents the first parameter.

[0134] In this embodiment, after the server-side model completes centralized training, the model update parameters on the device side can be obtained, and the model parameters are updated on the device side. The update strategy adopts: In federated learning or distributed learning scenarios, the model parameter update strategy is one of the cores, because it determines how to effectively use the collaboration between the central server and the edge device (or client) to optimize the model. The following is a detailed description of this strategy:

[0135] Parameter initialization: First, complete one or several rounds of centralized training on the server side to obtain the global model parameters r. These parameters usually represent the generalization ability on the entire dataset.

[0136] Personalized training on the device side: Next, the global model parameter r is sent to each edge device. Each device uses its own private data to train the model and obtains the model parameter r′ that adapts to the characteristics of the local data. This process can be regarded as fine-tuning on the device side, with the aim of making the model more consistent with the distribution of device-side data.

[0137] Parameter fusion: The key step of the update strategy is how to combine the global knowledge on the server side with the personalized knowledge on the device side. According to the formula you provided, the new model parameter r is obtained by linear interpolation of the global parameter r and the device parameter r′, specifically:

[0138]

[0139] Wherein, r″ represents the sixth parameter of the fourth model, a represents the update ratio, r represents the seventh parameter obtained by the server through training the model to be trained, and r′ represents the first parameter.

[0140] The a (update ratio) determines the degree of integration between the server-side model parameter r and the device-side model parameter r'. When a = 0.5, it means that both are given equal attention, neither completely favoring the universality of the global model nor over-emphasizing the particularity of the device-side data, achieving a balanced integration of the two. By adjusting the value of a, the influence of the global model on the personalized model can be flexibly controlled to adapt to different scenario requirements.

[0141] Feedback loop: The updated parameter r or its gradient on the device side can be uploaded to the server. The server aggregates and calculates new global model parameters again based on the update results of all participating devices. This cycle is repeated until the model converges or the stopping condition is met.

[0142] The update strategy enables the strategy to integrate the personalized needs of the device side while maintaining the generalization ability of the model, especially when the data distribution is very different, which can effectively improve the performance of the model on the edge device. The introduction of the update ratio a provides the flexibility to adjust the direction of the model update, which can be adjusted according to the needs of the actual application scenario, and realizes continuous adjustment from relying entirely on the server-side model to focusing on device-side personalization. Using the computing resources on the device side for personalized training reduces the computing burden of the server, which is particularly suitable for resource-constrained or data privacy-sensitive environments.

[0143] In short, this update strategy that combines server-side model parameters with device-side model parameters provides an efficient and flexible model optimization path for distributed learning systems by cleverly balancing global knowledge and personalized needs, which helps to improve the application effect of the model in diverse and personalized scenarios.

[0144] Finally, for step S107, in response to reaching the condition for stopping training, the third model and the fourth model are used as trained models.

[0145] When the model results converge, the training can be stopped, and the third model and the fourth model obtained in the current round of training can be used as the final model. The final model here can be a model that is spliced ​​together based on the number of layers initially divided by the third model and the fourth model.

[0146] It can be seen from the above embodiments that the model training method based on edge collaboration described in the embodiment of the present application splits the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model; based on the local data of the mobile device, the mobile device is used to perform knowledge distillation on the second model to obtain a distillation model; the first parameter of the distillation model is sent to the data processor by the mobile device to obtain the second parameter, and the second parameter is uploaded to the server by the data processor; the server is used to calculate the fourth parameter based on the second parameter and the third parameter of the model to be trained; based on the fourth parameter, the first model is updated by the server to obtain the third model and the fifth parameter of the third model; based on the fifth parameter, the distillation model is trained by the mobile device to obtain the sixth parameter of the fourth model; in response to reaching the condition of stopping training, the third model and the fourth model are used as trained models. The embodiment of the present application first constructs a large language model collaborative training framework based on federated learning and split learning. Under this framework, the participants do not need to exchange original data. Through the federated learning mechanism, the model parameters can be trained locally, and the encrypted model updates can be uploaded to the central server, ensuring data privacy and compliance. Taking into account the large number of parameter exchange processes in federated learning, an in-network aggregation method based on the data processor DPU is designed. The DPU is deployed in the switch to aggregate parameters while transmitting data, reduce network load, and increase overall throughput. At the same time, using the split learning strategy, the large language model is divided into smaller sub-models, which are deployed on different nodes for parallel training, and then the model updates are integrated through encryption protocols, which not only significantly improves the training speed, but also reduces the computational burden of individual participants. Secondly, the present invention designs a large language model fine-tuning method based on knowledge distillation. In the training of large language models, knowledge distillation can help this article transfer the knowledge of large models to small models, thereby realizing the lightweight and efficient model, and reducing the pressure of federated learning on the network transmission environment.

[0147] In some other embodiments, in order to verify the effectiveness of the present invention, the following experiments were conducted:

[0148] First, complete the model experiment of the control group: train the BERT-base model on the training task in the local hardware environment, record the model training time and the accuracy of the model in completing the task. Record the experimental results of independent BERT training on the local device:

[0149] Table 1. Model training results without splitting

[0150]

[0151] From the above experiments, it can be seen that as the batch size increases, the training time of a single iteration generally shows a downward trend. This is because when the batch size increases, the model can process more samples in one forward propagation and back propagation, making better use of hardware resources, thereby shortening the training time. According to experimental data, when the BERT model is trained for the binary sentiment classification task, appropriately increasing the batch size can effectively reduce the training time and improve the accuracy and F1 score of the model to a certain extent. However, when the batch size is too large, the improvement in accuracy and F1 score is not significant, and may even decrease slightly. In actual training, it is necessary to choose a moderate batch size to achieve a good balance between performance and efficiency.

[0152] Based on the control group experiment, the experiment of knowledge distillation training on the split model was carried out on the local device. The experimental data of migrating the large model to the small model is as follows:

[0153] Table 2. Split model knowledge distillation training results

[0154]

[0155] refer to Figure 6 , which is a schematic diagram for visualizing the knowledge distillation training results of the split model according to an embodiment of the present application.

[0156] from Figure 6 It can be seen that the increase in batch size from 1 to 10 corresponds to a gradual increase in the number of samples used in each training iteration. As the batch size increases from 1 to about 7, the accuracy (ac) of the model shows an upward trend, from 0.50450 to 0.86887, indicating that increasing the batch size within a certain range helps the student model to learn and converge better. However, when the batch size continues to increase to 8, 9, and 10, the accuracy does not continue to improve, and even decreases slightly when the batch size is 10. This may be because the excessively large batch size causes the training dynamics of the model to change and is no longer conducive to model optimization. On the other hand, as the batch size increases, the time required for a single iteration is also gradually increasing. This is because larger batches require more memory resources and may result in reduced hardware utilization, so the overall training time may increase.

[0157] In the federated aggregation simulation experiment, the computing resources of the device are limited, and only three clients are simulated, each with 1,000 pieces of configuration data and one server device.

[0158] Table 3 Average time records of federated learning clients

[0159]

[0160] Regarding congestion control, although its dynamic nature increases the complexity of the simulation, the behavioral model is set based on the standard TCP congestion control algorithm. In the early stages of the simulation, it can be simplified by assuming a basic congestion window growth and reduction rule, and then the model can be gradually refined as the simulation progresses. The actual transmission time calculation will be more complicated than the basic formula because it needs to consider more network factors, such as latency, packet loss, and the impact of congestion control mechanisms (such as TCP's slow start and congestion avoidance algorithms). A more comprehensive but simplified model can consider the combined effects of these factors, although accurate predictions are still challenging because network behavior can be highly dynamic and unpredictable. Considering bandwidth, number of packets, packet loss retransmission factor, and expected number of retransmitted packets, the expected network environment at the edge is set accordingly:

[0161] RTT (Round-Trip Time): The RTT in the actual network can range from tens of milliseconds to hundreds of milliseconds, or even higher. For local network simulation, set the RTT to 1-10ms.

[0162] Number of packets: The number of packets depends on the amount of data you want to transfer and the maximum segment size (MSS) of TCP. If you have 1MB of data to transfer and the MSS is 1460 bytes (excluding IP and TCP headers), you will need about 700 packets (taking into account the TCP header and possible fragmentation).

[0163] Packet loss retransmission factor: This factor is usually estimated based on historical data or network conditions. In a stable network, the packet loss rate may be very low. In simulation, an initial packet loss rate of 0.5% can be set.

[0164] Expected number of retransmitted packets: This number can be estimated by multiplying the packet loss rate by the total number of packets. For example, if the total number of packets is 700 and the packet loss rate is 0.5%, the expected number of retransmitted packets is about 3.5.

[0165] Transmission time = model size × 8 / bandwidth + (RTT × number of packets) + (packet loss retransmission factor × expected number of retransmission packets)

[0166] On the basis of completing the simulation experiments of each part, the data results obtained from the experiment are further analyzed from the following three aspects:

[0167] Table 4 Results analysis table

[0168]

[0169] (1) Classification accuracy analysis:

[0170] As can be seen from the table data, the prediction accuracy of the BERT-base model is 0.926926927, while the prediction accuracy of the S-Bert model is 0.86887. This shows that BERT-base performs significantly better than the S-Bert model in terms of accuracy. This means that BERT-base is a better choice in scenarios where tasks require high-precision predictions. However, if the application scenario is more sensitive to speed or resource consumption, it may be necessary to weigh accuracy and efficiency.

[0171] (2) Latency analysis (computational delay and communication delay):

[0172] Computational latency: From the perspective of single-batch training time, the training time of the BERT-base model is 185.603456 seconds, while the S-Bert model only takes 59.75997615 seconds, indicating that the S-Bert model has a significant advantage in training speed, and its training efficiency is about three times that of BERT-base. This shows that S-Bert has significantly reduced the computational complexity through some optimization methods (such as model pruning, quantization, etc.), thereby speeding up the training process.

[0173] Communication latency: Although the table does not directly provide communication latency data, changes in model size have a direct impact on communication latency. The transmission size of the S-Bert model is only 8,427KB, which is much smaller than BERT-base's 399,587KB. In distributed training or cloud deployment scenarios, smaller models mean faster model loading and update times, thereby reducing communication latency and improving the overall system response speed. In terms of network transmission time, the BERT-base model requires 3896.716 milliseconds, while the S-Bert model requires 1374.18 milliseconds. This means that S-Bert can complete network transmission faster when the model is deployed or updated, reducing the waiting time caused by model exchange, which is especially important for distributed training or scenarios with frequent model updates.

[0174] (3) Equipment load analysis: Model size changes

[0175] The significant reduction in model size (from 399,587KB for BERT-base to 8,427KB for S-Bert) has an important impact on device load. Smaller model size means lower memory usage. The S-Bert model occupies 3442MiB of memory during training, which is about 1700MiB less than 5144MiB for BERT-base, reducing the demand for hardware resources. This is especially important for devices with limited resources (such as mobile devices and edge computing nodes) because it allows complex models to be deployed on these platforms without causing resource overload or slow operation. Therefore, S-Bert shows obvious advantages in reducing the memory burden of devices and expanding the range of applicable devices.

[0176] In summary, the S-Bert model significantly improves training speed and reduces resource requirements at the expense of a certain degree of accuracy, making it the preferred solution for efficiency and resource-sensitive scenarios. BERT-base is more suitable for environments that have strict requirements on prediction accuracy and sufficient resources. The choice of model should be based on a comprehensive consideration of the accuracy, speed, and resource requirements of the specific application scenario.

[0177] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the described method.

[0178] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0179] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a model training device based on edge collaboration.

[0180] refer to Figure 7 , the model training device based on edge-end collaboration includes:

[0181] A splitting module 71 is configured to split the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model;

[0182] The mobile device 72 is configured to perform knowledge distillation on the second model based on local data of the mobile device to obtain a distilled model;

[0183] A mobile device 72 is configured to send the first parameter of the distillation model to a data processor to obtain a second parameter, and upload the second parameter to a server using the data processor;

[0184] The server 73 is configured to calculate a fourth parameter based on the second parameter and the third parameter of the model to be trained;

[0185] The server 73 is configured to obtain a third model and a fifth parameter of the third model by updating the first model based on the fourth parameter;

[0186] The mobile device 72 is configured to train the distillation model based on the fifth parameter to obtain a sixth parameter of the fourth model;

[0187] The stopping module 74 is configured to use the third model and the fourth model as trained models in response to the condition for stopping training being met.

[0188] For the convenience of description, the above device is described in terms of functions divided into various modules. Of course, when implementing the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0189] The device of the above embodiment is used to implement the corresponding model training method based on edge collaboration in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0190] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the model training method based on edge collaboration described in any of the above embodiments is implemented.

[0191] Figure 8 A more specific schematic diagram of the hardware structure of an electronic device provided in this embodiment is shown, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 in the device.

[0192] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0193] The memory 1020 may be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0194] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0195] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0196] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0197] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present specification, and does not necessarily include all the components shown in the figure.

[0198] The electronic device of the above-mentioned embodiment is used to implement the corresponding model training method based on edge collaboration in any of the above-mentioned embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0199] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the edge-to-edge collaboration-based model training method as described in any of the above embodiments.

[0200] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0201] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the model training method based on edge collaboration as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0202] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0203] In addition, to simplify the description and discussion, and in order not to make the embodiments of the present application difficult to understand, the known power supply / ground connection with the integrated circuit (IC) chip and other components may or may not be shown in the provided drawings. In addition, the device can be shown in the form of a block diagram to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform to be implemented in the embodiments of the present application (that is, these details should be fully within the scope of understanding of those skilled in the art). In the case of elaborating specific details (e.g., circuits) to describe exemplary embodiments of the present application, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0204] Although the present application has been described in conjunction with specific embodiments of the present application, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0205] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A model training method based on edge-end collaboration, characterized in that: include: Splitting the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model; Based on local data of the mobile device, using the mobile device to perform knowledge distillation on the second model to obtain a distilled model; Using a mobile device to send a first parameter of the distillation model to a data processor to obtain a second parameter, and using the data processor to upload the second parameter to a server; Using the server, based on the second parameter and the third parameter of the model to be trained, a fourth parameter is calculated; Based on the fourth parameter, using the server to update the first model to obtain a third model and a fifth parameter of the third model; Based on the fifth parameter, using the mobile device to train the distillation model to obtain a sixth parameter of the fourth model; In response to reaching a condition for stopping training, the third model and the fourth model are used as trained models.

2. The method according to claim 1, characterized in that The mobile device resource parameters include hardware resource parameters, delay sensitivity, load balancing, model structure, communication cost, security and model validity; The method of splitting the model to be trained according to the obtained mobile device resource parameters to obtain the first model and the second model includes: The model to be trained is split based on the hardware resource parameters, delay sensitivity, load balancing, model structure, communication cost, security and model effectiveness to obtain the first model and the second model.

3. The method according to claim 1, characterized in that The method of performing knowledge distillation on the second model based on the local data of the mobile device and using the mobile device to obtain a distilled model includes: Using the server to train the second model to obtain a fifth model; Based on the local data and the learning loss of the soft label distribution of the fifth model, the second model is subjected to knowledge distillation using the mobile device to obtain the distilled model.

4. The method according to claim 1, characterized in that The method of using a mobile device to send the first parameter of the distillation model to a data processor to obtain a second parameter, and using the data processor to upload the second parameter to a server includes: Sending the first parameter of the distillation model to the data processor using a mobile device; Using the data processor to perform aggregation processing on the first parameter to obtain the second parameter; The second parameter is uploaded to the server using the data processor.

5. The method according to claim 1, characterized in that The calculating a fourth parameter based on the third parameter of the model to be trained based on the second parameter comprises: Among them, w new Indicates the fourth parameter, w global Represents the third parameter, represents the second parameter, i represents the i-th switch, η represents the learning rate of the model to be trained, and N represents the number of mobile devices participating in this round of training.

6. The method according to claim 1, characterized in that The updating of the first model by using a server based on the fourth parameter to obtain a third model and a fifth parameter of the third model includes: The fourth parameter is used as a parameter of the first model, and the first model is updated by using a server to obtain a third model and a fifth parameter of the third model.

7. The method according to claim 1, characterized in that The sixth parameter is calculated by the following formula: Wherein, r″ represents the sixth parameter of the fourth model, a represents the update ratio, r represents the seventh parameter obtained by the server through training the model to be trained, and r′ represents the first parameter.

8. A model training device based on edge-end collaboration, characterized in that: include: A splitting module is configured to split the model to be trained according to the obtained mobile device resource parameters to obtain a first model and a second model; The mobile device is configured to perform knowledge distillation on the second model based on local data of the mobile device to obtain a distilled model; A mobile device configured to send the first parameter of the distillation model to a data processor to obtain a second parameter, and upload the second parameter to a server using the data processor; The server is configured to calculate a fourth parameter based on the second parameter and a third parameter of the model to be trained; The server is configured to obtain a third model and a fifth parameter of the third model by updating the first model based on the fourth parameter; A mobile device, configured to train the distillation model based on the fifth parameter to obtain a sixth parameter of the fourth model; The stopping module is configured to use the third model and the fourth model as trained models in response to reaching a condition for stopping training.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 7.

Images